Cost-effective land cover classification for remote sensing images

Dongwei Li, Shuliang Wang, Qiang He, Yun Yang

Introduction

Generating land cover maps is a critical task for urban planning, agricultural management, and environmental monitoring. However, as the volume of high-resolution satellite imagery grows, the computational cost of processing these datasets in the cloud has become a significant barrier. This paper addresses the financial burden of large-scale remote sensing classification by introducing an "early stopping" framework. By predicting when a clustering algorithm has reached a sufficient accuracy threshold, researchers can drastically reduce cloud computing costs without sacrificing the practical utility of the resulting maps.

The Long Tail Problem in Clustering

The core motivation for this research is the "long tail" phenomenon observed in unsupervised clustering algorithms, such as Fuzzy C-Means (FCM). As illustrated in Figure 1, these algorithms typically achieve the majority of their accuracy within the first few iterations. However, they continue to consume significant computational resources—often more than half of the total processing time—to achieve marginal gains in mathematical convergence. In many real-world applications, 100% mathematical convergence is unnecessary; an accuracy level of 85% to 95% is often considered a reliable benchmark for land cover classification.

The long tail phenomenon illustrates how accuracy plateaus, leading to diminishing returns on computational investment.

Proposed Methodology

The authors propose a generalized framework consisting of two distinct phases to optimize resource usage:

1. Training Phase

The system first runs the FCM algorithm on a sample of images to completion. It records the relationship between the "change rate of the objective function" ($\Delta \mathcal{J}$) and the "achieved accuracy," measured by the Rand Index. To make this relationship predictable, the authors use Local Outlier Factors (LOF) to detect and remove anomalies in the training data. They then train a Support Vector Regressor (SVR) to model the relationship between the objective function's behavior and the actual accuracy.

An example of calculating the Rand Index, the metric used to evaluate clustering accuracy between different partitions.

2. Testing/Production Phase

When processing new images, the system monitors the $\Delta \mathcal{J}$ in real-time. The pre-trained SVR model predicts the current accuracy level. Once the desired threshold (e.g., 90% or 95%) is reached, the algorithm stops immediately. The paper specifically advocates for Fuzzy C-Means over traditional K-Means because pixels in remote sensing often represent "mixed" land types (e.g., a pixel containing both road and vegetation), which fuzzy logic handles more effectively.

Experimental Results and Cost Savings

The study utilized the SpaceNet dataset, which contains over 17,533 high-resolution remote sensing images covering 67,000 square kilometers. The efficiency gains observed were substantial:

  • 85% Accuracy: Requires only 27.34% of the cost of 100% accuracy.
  • 95% Accuracy: Requires only 33.25% of the total cost.
  • 99.9% Accuracy: Still only requires 60.83% of the total cost.

To put this in financial terms: for a massive task like classifying the entire United States at 90% accuracy, the authors estimate their approach could save over $1,593,490.18 in cloud fees per single-use.

Conclusion

By leveraging SVR to predict accuracy thresholds and implementing early stopping, researchers can drastically reduce the cost of remote sensing image classification. While the results are promising, future work will focus on testing the framework across a wider variety of sensors and geographic regions to ensure the SVR model generalizes well across different types of terrain and satellite hardware.

Summary updated

Abstract

Land cover maps are of vital importance to various fields such as land use policy development, ecosystem services, urban planning and agriculture monitoring, which are mainly generated from remote sensing image classification techniques. Traditional land cover classification usually needs tremendous computational resources, which often becomes a huge burden to the remote sensing community. Undoubtedly cloud computing is one of the best choices for land cover classification, however, if not managed properly, the computation cost on the cloud could be surprisingly high. Recently, cutting the unnecessary computation long tail has become a promising solution for saving cost in the cloud. For land cover classification, it is generally not necessary to achieve the best accuracy and 85% can be regarded as a reliable land cover classification. Therefore, in this paper, we propose a framework for cost-effective remote sensing classification. Given the desired accuracy, the clustering algorithm can stop early for cost-saving whilst achieving sufficient accuracy for land cover image classification. Experimental results show that achieving 85%-99.9% accuracy needs only 27.34%-60.83% of the total cloud computation cost for achieving a 100% accuracy. To put it into perspective, for the US land cover classification example, the proposed approach can save over $1,593,490.18 for the government in each single-use when the desired accuracy is 90%.
BibTeX
@article{li2022cost,
  author = {Dongwei Li and Shuliang Wang and Qiang He and Yun Yang},
  title = {{Cost-effective land cover classification for remote sensing images}},
  journal = {J. Cloud Comput.},
  year = {2022},
  doi = {10.1186/s13677-022-00335-0},
}
Figure